Google Launches Gemini Spark: A 24/7 Personal AI Agent That Keeps Working After You Log Off

Google opened its I/O 2026 keynote Tuesday with Gemini Spark, a new agentic assistant that runs continuously in the cloud, plugs into Gmail and Workspace, and ships with launch-day connectors to Canva, OpenTable, and Instacart. It is the most ambitious personal AI agent any of the major labs has released to consumers, and it ships in beta to US AI Ultra subscribers next week.

The pitch is direct: Spark keeps working after you close your laptop or lock your phone. It is not a chat assistant you have to babysit — it is a background process you assign tasks to.

What Spark Actually Does

Built on Gemini 3.5 Flash and an agentic harness called Google Antigravity, Spark runs as a cloud service rather than an on-device tool. That distinction is the whole product. Because it lives in Google's data centers, it can keep executing multi-step workflows even when the user is offline.

The launch demos centered on three categories. Inbox management: Spark scans Gmail and produces a daily list of items that actually require your attention, then writes the first-pass replies. Recurring tasks: it can be programmed to flag hidden fees on monthly credit card statements, or pull deadline updates from long email threads. Cross-app workflows: a single instruction can string together Docs, Slides, Calendar, and a third-party tool like Canva to produce a finished deliverable.

The third-party connector list at launch is small — Canva, OpenTable, and Instacart — but Google says it is built on Model Context Protocol (MCP), the same open standard Anthropic introduced last year and that most major labs have now adopted. More partners are expected in the coming weeks.

Why this matters: until today, "personal AI agent" was mostly a demo category. ChatGPT's Operator is browser-driven and stops when the tab closes. Claude's Cowork agents are powerful but require you to keep the session open. Spark is the first mass-market agent that genuinely runs without supervision.

The Halo Layer on Android

Spark is not just a Gemini app feature. Google also previewed Android Halo, a new system-level surface that lets users monitor what their agent is doing from any device. From an Android phone, you can see Spark's queue, intervene on a specific task, or hand a task off to your laptop mid-execution.

Halo is the more interesting design choice. Most agentic interfaces are buried inside a single app. By promoting agent oversight to an OS-level layer, Google is treating long-running agents the way operating systems treat background processes — visible, pausable, and persistent across sessions.

For Mac users, Google is bringing Spark into the Gemini desktop app later this summer with local file access. That puts it in direct competition with Anthropic's Claude Cowork, which has shipped local file capabilities since late 2025.

Industry Impact

Spark is the most aggressive consumer agent launch from a hyperscaler to date, and it lands in a market where consumer agentic AI was still considered "two years out" as recently as January.

For OpenAI, Spark resets the bar for what a paid ChatGPT tier needs to offer. ChatGPT Pro and Operator remain strong on browser automation, but neither runs continuously, and neither has Google's distribution. The Gemini app reportedly crossed 900 million active users this quarter, giving Spark a deployment surface no rival can match.

For Anthropic, the pressure is more strategic. Claude's enterprise agent business has been the company's growth story for 18 months. Spark does not directly target enterprises, but the consumer template — always-on, MCP-connected, OS-integrated — is the same template Anthropic will need to ship for its own consumer push.

For SaaS vendors, MCP connectors are now the price of admission. Canva, OpenTable, and Instacart got launch-day placement on the most visible AI surface in consumer software. Every category leader will now be calculating whether to ship their own MCP server or risk being routed around.

Expert Perspectives

Reaction on X has split predictably. Optimists point to the integration depth — Gmail, Calendar, Docs, Sheets, plus three third-party brands — as something only Google could ship on day one. Pessimists raise the obvious privacy question: an agent that runs 24/7 inside your Gmail is also an agent that reads your Gmail 24/7.

Google's response is that Spark operates under user-defined instructions and that all actions show up in a visible activity log. Whether that's enough to satisfy regulators, especially in the EU, is the open question. The DMA already restricts Google's data-sharing across services, and a cross-service agent sits awkwardly inside that regime.

Several researchers also flagged a model question: Spark runs on Gemini 3.5 Flash, not the more capable Gemini 3.1 Pro. The choice is almost certainly about cost — Flash is 40% cheaper on input than Pro — but it caps how complex an autonomous task Spark can reliably handle today.

What to Watch

Spark rolls out to trusted testers this week and lands in beta for US AI Ultra subscribers next week. Three things to watch over the next 30 days. First, the next wave of MCP connectors — calendar, banking, and travel are the obvious gaps. Second, the macOS rollout this summer and whether Halo-style oversight extends beyond Android. Third, the reliability data: 24/7 agents fail in new and embarrassing ways, and the first viral Spark-mistake screenshot will shape consumer trust.

The Bottom Line

Gemini Spark is the most credible mass-market AI agent launched so far, and the first one that treats long-running autonomy as a default rather than a feature flag. If it works as demoed, the post-chat era of consumer AI starts at I/O 2026. If it stumbles in the wild, every other lab gets a clearer map of which failure modes to avoid.